Abstract: A technique for fine-granularity speed binning for a processing device is provided. The processing device includes a plurality of clock domains, each of which may be clocked with independent clock signals. The clock frequency at which a particular clock domain may operate is determined based on the longest propagation delay between clocked elements in that particular clock domain. The processing device includes measurement circuits for each clock domain that measure such propagation delay. The measurement circuits are replica propagation delay paths of actual circuit elements within each particular clock domain. A speed bin for each clock domain is determined based on the propagation delay measured for the measurement circuits for a particular clock domain. Specifically, a speed bin is chosen that is associated with the fastest clock speed whose clock period is longer than the slowest propagation delay measured for the measurement circuit for the clock domain.

Abstract: Various energy efficient data encoding schemes and computing devices are disclosed. In one aspect, a method of transmitting data from a transmitter to a receiver connected by plural wires is provided. The method includes sending from the transmitter on at least one but not all of the wires a first wave form that has first and second signal transitions. The receiver receives the first waveform and measures a first duration between the first and second signal transitions using a locally generated clock signal not received from the transmitter. The first duration is indicative of a first particular data value.

Abstract: A method of and device for transferring data is provided. The method includes determining a difference between a data segment that was transferred last relative to each of one or more data segments available to be transferred next. In some embodiments, for so long as no data segment available to be sent has been waiting too long, the data segment chosen to be sent next is the data segment having the smallest difference relative to the data segment transferred last. The chosen data segment is then transmitted as the next data segment transferred.

Abstract: A method and apparatus of asynchronous scheduling in a graphics device includes sending one or more instructions from an instruction scheduler to one or more instruction first-in/first-out (FIFO) devices. An instruction in the one or more FIFO devices is selected for execution by a single-instruction/multiple-data (SIMD) pipeline unit. It is determined whether all operands for the selected instruction are available for execution of the instruction, and if all the operands are available, the selected instruction is executed on the SIMD pipeline unit. The self-timed arithmetic pipeline unit (SIMD pipeline unit) is effectively encapsulated in a synchronous, (e.g., clocked by global clock), scheduler and register file environment.

Abstract: An asynchronous pipeline includes a first stage and one or more second stages. A controller provides control signals to the first stage to indicate a modification to an operating speed of the first stage. The modification is determined based on a comparison of a completion status of the first stage to one or more completion statuses of the one or more second stages. In some cases, the controller provides control signals indicating modifications to an operating voltage applied to the first stage and a drive strength of a buffer in the first stage. Modules can be used to determine the completion statuses of the first stage and the one or more second stages based on the monitored output signals generated by the stages, output signals from replica critical paths associated with the stages, or a lookup table that indicates estimated completion times.

Abstract: Systems, apparatuses, and methods for optimizing neural network training with a first-in, last-out (FILO) buffer are disclosed. A processor executes a training run of a neural network implementation by performing multiple passes and adjusting weights of the neural network layers on each pass. Each training phase includes a forward pass and a backward pass. During the forward pass, each layer, in order from first layer to last layer, stores its weights in the FILO buffer. An error is calculated for the neural network at the end of the forward pass. Then, during the backward pass, each layer, in order from last layer to first layer, retrieves the corresponding weights from the FILO buffer. Gradients are calculated based on the error so as to update the weights of the layer for the next pass through the neural network.

Abstract: A method of and device for removing a processor from a low power mode. The method includes and the device provides for performing multiple processor start-up tasks in parallel. Memory interface training between the processor and memory and restoration and initialization of the processor are performed in parallel with each other and with a serial bus controller entering serial bus training to facilitate communication between the processor and a system controller.

Abstract: Methods and apparatus are described. A method, implemented in a decoder, includes receiving two or more signals from an encoder over two or more respective wires. At least one of the two or more signals includes at least one code that was recoded by the encoder. The decoder receives a recoding table. The recoding table provides a mapping indicating the recoding for each code that was recoded by the encoder in the received two or more signals. The decoder decodes the two or more received signals using the received recoding table.

Abstract: A method, computer program product, and system are provided for associating one or more memory buffers in a computing system with a plurality of memory channels. The method can include associating a first memory buffer to a first plurality of memory banks, where the first plurality of memory banks spans over a first set of one or more memory channels. Similarly, the method can include associating a second memory buffer to a second plurality of memory banks, where the second plurality of memory banks spans over a second set of one or more memory channels. The method can also include associating a first sequence identifier and a second sequence identifier with the first memory buffer and the second memory buffer, respectively. Further, the method can include accessing the first and second memory buffers based on the first and second sequence identifiers.

Abstract: An asynchronous pipeline includes a first stage and one or more second stages. A controller provides control signals to the first stage to indicate a modification to an operating speed of the first stage. The modification is determined based on a comparison of a completion status of the first stage to one or more completion statuses of the one or more second stages. In some cases, the controller provides control signals indicating modifications to an operating voltage applied to the first stage and a drive strength of a buffer in the first stage. Modules can be used to determine the completion statuses of the first stage and the one or more second stages based on the monitored output signals generated by the stages, output signals from replica critical paths associated with the stages, or a lookup table that indicates estimated completion times.

Abstract: One or more processors are operative to carry out neural network operations and include a plurality of compute units (CUs) configurable for neural network operations. The neural network compute unit remapping logic detects a condition to remap neural network compute units that are currently used in carrying out a neural network operation in a processor with at least one replacement compute unit that is not currently being used to carry out the neural network operation. In response to detecting the condition, the logic remaps a logical address of at least one currently used compute unit to a different physical address that corresponds to the replacement compute unit and causes the replacement compute unit to carry out neural network operations.

Abstract: Various energy efficient data encoding schemes and computing devices are disclosed. In one aspect, a method of transmitting data from a transmitter to a receiver connected by plural wires is provided. The method includes sending from the transmitter on at least one but not all of the wires a first wave form that has first and second signal transitions. The receiver receives the first waveform and measures a first duration between the first and second signal transitions using a locally generated clock signal not received from the transmitter. The first duration is indicative of a first particular data value.

Abstract: Control of power supplied to a machine intelligence (MI) processor is provided with an energy reservoir and power switching circuitry coupled to a power supply, the energy reservoir, and to power delivery circuitry of the MI processor. Control circuitry directs the power switching circuitry to charge the energy reservoir from the power supply or discharge the energy reservoir to the MI processor based on MI state information obtained from the MI processor. Processes for charging and discharging such an energy reservoir are provided. Processes for analyzing state information of the MI processor and configuring the control circuitry are also provided.

Abstract: A sampling circuit automatically resamples the data from another timing domain until the sampled data is represented correctly in the new domain by assuring that no metastable states exist. If a metastable state exists, a sampling signal recirculates through the sampling circuit until the metastable state no longer exists. A comparison of input data to sampled data is used to determine the existence of a metastable state.

Abstract: A processing unit includes a plurality of components configured to execute instructions and a controller. The controller is configured to determine a power consumption of the processing unit, determine a waiting status of the processing unit based on waiting statuses of components, and selectively modify an operating state of the processing unit based on the waiting status and the power consumption of the processing unit. In some cases, the operating state is modified in response to a percentage of the components that are waiting for an action to complete being below a threshold percentage and the power consumption of the processing unit being below a power limit. In some cases, the controller identifies a pattern in the power consumption by the processing unit and modifies the operating state of the processing unit to increase the power consumption of the processing unit based on the pattern identified by the controller.

Abstract: A conversion unit converts operands from a conventional number system that represents each binary number in the operands as one bit to redundant number system (RNS) operands that represent each binary number as a plurality of bits. An arithmetic logic unit performs an arithmetic operation on the RNS operands in a direction from a most significant bit (MSB) to a least significant bit (LSB). The arithmetic logic unit stops performing the arithmetic operation prior to performing the arithmetic operation on a target binary number indicated by a dynamic precision associated with the RNS operands. In some cases, a power supply provides power to bit slices in the arithmetic logic unit and a clock signal generator provides clock signals to the bit slices. Gate logic is configured to gate the power or the clock signals provided to a subset of the bit slices.

Abstract: Systems, apparatuses, and methods for managing power consumption for a neural network implemented on multiple graphics processing units (GPUs) are disclosed. A computing system includes a plurality of GPUs implementing a neural network. In one implementation, the plurality of GPUs draw power from a common power supply. To prevent the power consumption of the system from exceeding a power limit for long durations, the GPUs coordinate the scheduling of tasks of the neural network. At least one or more first GPUs schedule their computation tasks so as not to overlap with the computation tasks of one or more second GPUs. In this way, the system spends less time consuming power in excess of a power limit, allowing the neural network to be implemented in a more power efficient manner.

Abstract: An apparatus includes a plurality of registers to store sets of state information that represent a state history of a processing unit. The apparatus also includes a power management advisor (PMA) to generate a signal based on the sets of state information, wherein the signal indicates a probability that a power state transition of the processing unit achieves a target outcome. In some cases, the signal is provided to a power management controller including hardware circuitry that initiates a power state transition of the processing unit based on the signal and inputs to the power management controller that represent a subset of the state information corresponding to a current power state of the processing unit.

Abstract: Various energy efficient data encoding schemes and computing devices are disclosed. In one aspect, a method of transmitting data from a transmitter to a receiver connected by plural wires is provided. The method includes sending from the transmitter on at least one but not all of the wires a first wave form that has first and second signal transitions. The receiver receives the first waveform and measures a first duration between the first and second signal transitions using a locally generated clock signal not received from the transmitter. The first duration is indicative of a first particular data value.

Abstract: Systems, apparatuses, and methods for reducing the toggle rates on buses are disclosed. A computing system includes a source which provides packets for transmission on a bus. The packet is compressed by a compression engine. The compressed data format of the packet includes locations (bit positions) referred to as holes which do not include valid data. A bus configuration module identifies the locations of the holes and replaces the holes with information from a previous packet transmitted earlier on the bus. The bus configuration module also determines a new transmission bus width for the packet for lowering the bus toggle rate on the bus during transmission.